ECMWF opens ten-day AI pollution forecasts to public view

Public charts now show an experimental AI forecast of atmospheric pollution up to ten days ahead. ECMWF announced the release on 16 October, bringing AIFS-COMPO alongside the operational Copernicus Atmosphere Monitoring Service system, which issues five-day global forecasts twice daily.[3]

The maps cover particles and gases, including dust, wildfire-related aerosols, fine particulate matter, nitrogen dioxide and ozone. Products show conditions at the surface, pressure levels or across an atmospheric column. ECMWF's announcement links to the charts; the underlying forecast data will be released later as the system matures.[1]

A forecast in fifty seconds

The preprint reports a five-day global forecast in 50 seconds on one GPU, versus roughly 1,000 seconds on 8,000 CPUs for IFS-COMPO. Those timings concern different computing architectures and resolutions, not measured electricity savings.[2]

The speed gives researchers room to extend the forecast horizon. But following pollution requires more than predicting wind: transport, emissions, chemical reactions and meteorology all shape how concentrations evolve. IFS-COMPO calculates these processes using physical and chemical equations.[3]

AIFS-COMPO learns the evolution of weather and composition together from CAMS reanalysis, analysis and forecast records. Its graph-transformer architecture advances predictions in three-hour steps. The research team's EGU account describes training first on reanalysis and then adapting the model using operational analysis and forecast data.[4]

The live starting point still comes from the operational system, which assimilates satellite measurements to estimate the atmosphere's current state. AIFS-COMPO forecasts forward from that analysis. The observing network and the operational analysis remain essential to supplying the initial conditions.[1]

Checked against the air

Testing used independent observations from 2024. For PM2.5 and PM10 across North America, Europe and China, AI root mean square error was lower in four of six comparisons. The other two—PM2.5 in Europe and PM10 in North America—were comparable with IFS-COMPO.[2]

These are overlapping particle-size categories, covering diameters up to approximately 2.5 and 10 micrometres. Root mean square error penalises larger misses more strongly; lower values mean forecasts stayed closer to measured concentrations under that comparison.[2]

The EGU account also reports lower aerosol-optical-depth error against observations. This quantity describes particles' effect on light passing through the atmosphere. It provides another way to test a forecast besides measurements of pollution at ground level.[4]

In a separate August–December 2024 case study, ten-day AI forecasts captured the Antarctic ozone hole's development and extent. The AI's roughly 80-kilometre grid was coarser than the physical model's approximately 40-kilometre grid in the evaluation.[2]

That scale is useful for following regional and global patterns. It cannot resolve the difference between neighbouring streets. A ten-day horizon also describes how far the model runs, not a promise of identical accuracy for every pollutant, place and lead time.

Giving emissions their own input

The experimental model predicts fewer chemical species than IFS-COMPO and omits deposition fluxes. It also has no explicit input for changing emissions. ECMWF is working to add fire-emission information from the CAMS Global Fire Assimilation System and other emissions datasets, enabling different emissions scenarios.[1]

That would add a useful question to where pollution is heading: how would its path change if the emissions entering the atmosphere changed? The newly public charts extend the view forward. Explicit emissions inputs would let researchers vary what enters the air throughout a forecast.

Report an error or suggest a correction